RiO-DETRv2: Geometry-Aware Uncertainty Modeling Meets Real-Time Oriented Object Detection Transformer

Introduction

• Oriented detection needs precise boxes for rotated objects, but deterministic box parameterizations limit localization gains while preserving real-time efficiency.The passage notes applications in aerial imagery, remote sensing, and scene text understanding.Source: Introduction · p. 1
• Distributional regression represents boundary locations as probability distributions rather than single scalars, providing richer supervision and progressive localization refinement.Its extension to oriented detection faces distinct geometric challenges.Source: Introduction · p. 1
• Figure 1 contrasts localization errors that pass IoU 0.5 but fail 0.75 with detector comparisons of AP50-95 and AP50 against inference latency on DOTA-1.0.The detector comparisons use single-scale training and testing.Source: Figure 1 · p. 1
• Global-axis boundary targets change with object rotation, angular refinement must move from global search to local estimation, and separate branch criteria may misalign with complete-OBB quality.These issues can entangle boundary uncertainty with orientation or yield high-confidence predictions with unreliable overall localization.Source: Introduction · pp. 1–2
• RiO-DETRv2 combines object-aligned boundary distributions, circular-to-local angular residual refinement, and complete-OBB quality-aligned learning.The framework models spatial and angular uncertainty according to their geometric structures while aligning optimization at the complete-OBB level.Source: Introduction · p. 2
• 54.1 mAP50-95 on DOTA-1.0 is achieved by RiO-DETRv2-x; across model scales and schedules, it improves accuracy over RiO-DETR with comparable computational cost.The reported improvements are particularly pronounced on high-IoU-threshold metrics.Source: Introduction · p. 2

CNN-based Oriented Object Detection

• CNN-based oriented detectors adapt horizontal frameworks through rotated proposals, aligned RoI features, or orientation heads; many use dense predictions, NMS, and rotation-specific modules.Two-stage examples include RoI Transformer and Oriented R-CNN; one-stage examples include S2ANet and FCOS-based variants.Source: CNN-based Oriented Object Detection · p. 2

DETR-based Real-Time and Oriented Detection

• Real-time DETRs improve efficiency through encoding, query initialization, distributional refinement, and supervision; oriented DETRs add rotation-aware proposals, alignment, matching, or periodic refinement.The passage identifies RiO-DETR as a real-time oriented DETR, while noting that these methods largely retain deterministic OBB parameterizations.Source: DETR-based Real-Time and Oriented… · p. 2

Angle Representation and Distributional Modeling

• Angle methods address periodicity through circular classification, coded labels, or phase representations, while GWD and KLD optimize distances between OBB Gaussian representations.These approaches typically decode a deterministic OBB rather than progressively refining distributions for object-aligned boundaries and periodic angular residuals.Source: Angle Representation and… · p. 2

Object-Aligned Boundary Distribution Refinement

• Boundary distributions use a canonical object-aligned frame, preserving consistent physical side meanings across rotations and separating boundary geometry from orientation estimation.Training defines side-distance residuals in the target-aligned frame; inference decodes local geometry using the predicted orientation.Source: Object-Aligned Boundary Distribution… · p. 3, Figure 3 · p. 4
• Side-specific N-bin logits accumulate additive refinements across decoder layers, then decode into continuous normalized boundary residuals.The object frame preserves side-wise semantics, enabling progressive refinement.Source: Object-Aligned Boundary Distribution… · p. 3

Circular-to-Local Angular Residual Refinement

• Angular refinement begins with a global circular search, then contracts its support across layers for progressively finer local orientation updates.The first-layer circular first moment handles the periodic boundary; later targets are clipped to the current local support.Source: Circular-to-Local Angular Residual… · p. 4
• Unified OBB Quality-Aligned Learning uses complete-box localization quality to coordinate classification and both boundary and angular refinement objectives.Classification confidence is also calibrated using uncertainty statistics from boundary and angular distributions.Source: Unified OBB Quality-Aligned Learning · p. 4, Distribution-aware confidence… · pp. 4–5

Experiments

• Experiments cover DOTA-1.0, DIOR-R, and AI-TOD-R using official train/test protocols; latency is measured on one NVIDIA T4 with TensorRT 10 and FP16.DOTA-1.0 uses 1024 × 1024 crops at stride 824; DIOR-R contains 23,463 images, and AI-TOD-R contains 28,036.Source: Experiments · p. 5, DOTA-1.0 · p. 5, DIOR-R · p. 5

Visualization Analysis

• Figure 4 compares initial and final OBBs above layer-wise distributions; boundary predictions concentrate in the object-aligned frame as angular corrections shift from large periodic to fine local adjustments.The visualization analysis describes a coarse-to-fine angular residual pattern across decoder layers.Source: Figure 4 · p. 6, Visualization Analysis · p. 7

Comparison with State-of-the-Art Methods

• RiO-DETRv2-n/s/m/l/x achieves 49.3/51.8/53.2/53.7/54.1 mAP50-95 on DOTA-1.0, beating corresponding baselines by 2.2/2.4/1.6/1.2/1.5 points with marginal parameter and FLOP increases.AP75 gains further indicate improved fine-grained OBB localization.Source: Results on DOTA-1.0 · p. 6
• RiO-DETRv2-x reaches 53.46 mAP50-95 and 78.32 AP50 on DIOR-R, outperforming YOLO26x-obb by 1.18 and 1.84 points at similar computational cost and latency.Improvements are more evident at AP75, under stricter IoU requirements.Source: Results on DIOR-R · pp. 6–7, Table 2 · p. 6
• On AI-TOD-R, RiO-DETRv2-m raises mAP50-95 from 18.2 to 21.0 and AP75 from 6.2 to 10.4 over RiO-DETR.RiO-DETRv2-x gains 2.0 mAP50-95 and 4.0 AP50; it also outperforms YOLO26-obb in mAP50-95 at both scales under comparable budgets.Source: Table 3 · p. 6, Results on AI-TOD-R · p. 7

Ablation Studies

• Combining OABDR and CTLAR raises AP50-95 from 51.6 to 52.7 on DOTA-1.0 and from 48.0 to 51.2 on DIOR-R; UniOBB-QAL further raises it to 53.2 and 52.6, respectively.The authors attribute a slight DOTA-1.0 AP50 decrease to confidence calibration that reduces confidence for lower-quality complete OBBs.Source: Ablation on Main Components · p. 7
• On DOTA-1.0, object-aligned distributions reach 52.88 AP50-95 and 54.66 AP75, versus 52.19 AP50-95 for global-axis distributions and 51.80 for RiO-DETR.The reported gain is attributed to geometry-consistent boundaries preserving stable side semantics under arbitrary rotations.Source: Ablation on OABDR · p. 7
• Circular-to-local angular refinement reaches 53.16 AP50-95, 78.39 AP50, and 55.13 AP75, improving on circular decoding alone by 0.14, 0.25, and 0.35 points, respectively.The ablation contrasts periodicity-aware circular modeling with linear angular distributions and progressive local refinement.Source: Ablation on CTLAR · p. 7, Table 6 · p. 7

• The framework models boundaries in an object-aligned frame, refines orientation through circular-to-local residual distributions, and aligns classification and localization using complete-OBB quality.Source: Conclusion · p. 7
• Experiments on DOTA-1.0, DIOR-R, and AI-TOD-R show consistent localization gains with comparable computational cost and a better speed–accuracy trade-off.Source: Conclusion · p. 7
• The authors observe a growing recall–precision trade-off as detector performance improves and encourage comprehensive threshold-based metrics alongside conventional metrics.Source: Conclusion · p. 7

A.1 Boundary Target Assignment

• Adjacent bins on the non-uniform candidate set receive interpolation weights for boundary-distribution supervision.The target is a normalized side residual in the object-aligned frame.Source: §A.1 · p. 10

A.2 Inverse Normalization

• The inverse-normalization step uses reference side distances, normalization factors, and a regression scale to recover refined side distances.The excerpt identifies these quantities but does not state their full recovery expression.Source: §A.2 · p. 10

A.3 Layer-wise OBB Decoding

• Refined side distances are decoded in the canonical frame and transformed back using the layer’s predicted orientation.This describes the layer-wise conversion from local geometry to an OBB.Source: §A.3 · p. 10

B Additional Details of CTLAR

• Angular refinement uses a residual to the matched target orientation, with full-domain circular interpolation initially and clipped local supports in later layers.The supplied text identifies the residual and the change in support across layers.Source: B Additional Details of CTLAR · p. 10

C Distribution-aware Confidence Calibration

• Boundary concentration statistics use top-K side-distribution probabilities, while angular calibration uses entropy and final-box anisotropy descriptors; a lightweight MLP calibrates the classification logit.The boundary and angular descriptors are used for distribution-aware confidence calibration.Source: C Distribution-aware Confidence… · p. 10

D.1 Method-specific Hyperparameters

• For AI-TOD-R, all models use 900 queries instead of the default 300.Table 7 summarizes the hyperparameters introduced by RiO-DETRv2.Source: §D.1 · p. 10

D.2 Latency Measurement

• Latency measures the complete GPU inference pipeline from network input to final detection outputs, using a single NVIDIA T4 GPU and batch size one.The measurements are for single-image end-to-end inference.Source: §D.2 · p. 10
• Dynamic-shape engines are benchmarked at optimization profile 0’s OPT shape, with input and output GPU buffers allocated once.TensorRT engine execution uses asynchronous execution APIs selected by engine version.Source: §D.2 · p. 10
• Rotated NMS is included for detectors that require post-processing, while end-to-end detectors use no external detection post-processing.The protocol measures each detector according to its actual inference pipeline.Source: §D.2 · p. 10
• Dataset-level operations such as DOTA patch creation and prediction merging, image loading, host-to-device transfer, and other CPU preprocessing are excluded.The reported scope is a single input image or image patch.Source: §D.2 · p. 10
• Timing follows 100 asynchronous warm-up iterations and averages 300 back-to-back inference runs measured with CUDA events.The start and end events are recorded on the same CUDA stream, followed by synchronization.Source: §D.2 · pp. 10–11

Aspect-ratio analysis

• For near-square objects, standard angular error develops a secondary mode near 90° despite nearly unchanged matched-box IoU, consistent with canonical-axis exchange.Angular distribution entropy increases toward the square limit.Source: Figure 5 · p. 12, Table 9 · p. 13

Distribution-error correspondence

• Boundary entropy correlates 0.782 with complete-OBB localization error, while mean boundary peak probability correlates −0.763 with 1−IoUOBB.These are Spearman correlations reported for matched DIOR-R detections.Source: E Empirical Analysis of Distributional… · p. 12, Table 8 · p. 12
• Mean and final-layer angular entropy correlate 0.672 and 0.685 with angular error; one minus final angular peak probability correlates 0.684.Angular statistics are evaluated on objects with aspect ratio at least 1.5 because near-square OBB orientation is intrinsically ambiguous.Source: E Empirical Analysis of Distributional… · pp. 11–12

Error-retention analysis

• Retaining the 10% lowest-uncertainty matched detections reduces mean 1 −IoUOBB from 0.180 to 0.049 (72.5%) and mean angular error from 1.877° to 0.240° (87.2%).Both retention curves remain substantially below random retention.Source: E Empirical Analysis of Distributional… · p. 12, Figure 7 · p. 14

Probability assigned to the regression target

• Boundary and angular distributions assign realized regression targets lower NLL than uniform references: 2.434 versus log(33) = 3.497, and 1.677 versus log(36) = 3.584.The corresponding top-1-bin calibration errors are 0.0057 and 0.0208, respectively.Source: E Empirical Analysis of Distributional… · p. 12

Calibration over all retained predictions

• The angle-aware estimator reduces ECE50 by 7.4%, NLL50 by 6.5%, Brier score by 6.4%, and Q-ECE by 6.7% versus the mean of two boundary-only runs.At τ = 0.75, ECE is 0.0521 versus 0.0556–0.0561, and NLL is 0.0777 versus 0.0812–0.0816.Source: E Empirical Analysis of Distributional… · p. 13, Table 10 · p. 14
• Figure 6 shows competitive selective risk across the complete coverage range, so the calibration improvement does not compromise confidence ordering.The estimator is compared with the mean of the two boundary-only runs.Source: E Empirical Analysis of Distributional… · p. 13, Figure 6 · p. 13

Discussion

• Across the analyses, uncertainty statistics correlate with geometric error, rank reliability, and support confidence calibration against complete OBB quality.The study evaluates pretrained RiO-DETRv2-x on the complete DIOR-R test set.Source: E Empirical Analysis of Distributional… · pp. 11, 13, 14

F Analysis of Near-Square Objects

• The analysis uses RiO-DETRv2-x predictions on DIOR-R’s 11,738-image, 124,438-object test set, grouping ground-truth boxes by aspect ratio.Predictions are matched class-wise in descending confidence order at IoU thresholds 0.5 and 0.75; groups are [1.0, 1.2), [1.2, 1.5), and [1.5, ∞).Source: F Analysis of Near-Square Objects · p. 14
• For true positives at IoU 0.5, the analysis reports standard π-periodic angular error and separately counts axis switches when E180 > 45° and Eaxis < 10°.Angular entropy is the normalized distribution entropy averaged over decoder layers.Source: F Analysis of Near-Square Objects · p. 14
• 25.71° angular MAE and a 20.69% axis-switch rate in the 1.0 ≤ r < 1.2 group contrast with an axis-equivalent angular MAE of 5.29°.A secondary mode near 90° further supports canonical-axis exchange as the source of much of the standard angular error.Source: F Analysis of Near-Square Objects · p. 14
• Across all three aspect-ratio groups, mean matched IoU varies by only 0.008 and R@75 remains between 71.30 and 72.68, while angular entropy rises from 0.380 to 0.515.The results indicate that near-square ambiguity is largely parameterization-level, preserves OBB geometry, and is reflected by higher angular uncertainty.Source: F Analysis of Near-Square Objects · p. 14